Building Prompts in AI Hub
Create a prompt from a template or from scratch, give it inputs and knowledge, then tune the model and settings so it behaves the way your business needs.
The thing you build once and use everywhere
Nadia keeps writing the same instruction. In one flow she asks a model to summarise a delivery complaint. In an app she asks it again, slightly differently. In a third place a colleague wrote their own version and it returns a different shape of answer.
A prompt in AI hub fixes that. You build the instruction once, test it, save it, and then call it from a flow, an app, or an agent. One definition, many callers.
A prompt is a saved instruction for an AI model. You write it carefully once — “read this complaint and give me three bullet points” — and then any app or flow can use it. If you improve the wording later, everything that calls it gets the better version.
Before you can build one
Prompts have real prerequisites, and the exam likes prerequisites because they are the difference between a feature that appears and a feature that does not.
What has to be true first
- The environment must have Dataverse installed. No Dataverse, no prompts.
- The environment must be in a supported region. Availability differs by region and changes over time.
- The tenant needs credits — prompts are a metered capability, not a free one.
If a scenario says the Prompts area is missing or a prompt cannot be created, walk that list before you blame permissions.
Two ways to start
| Starting point | What you get | Best when |
|---|---|---|
| Prompt template | A prewritten prompt for a common task — summarise, classify, extract, translate, analyse sentiment | The task is a standard one and you want a working baseline in seconds |
| Build your own | An empty canvas where you write the instruction and define every input yourself | The task is specific to your business, or no template is close enough |
Templates are not a lesser option. Starting from one and rewriting the instruction is a perfectly normal path — you keep the input wiring and replace the words.
In Power Apps or Power Automate you reach both from AI hub → Prompts, then either pick a template or choose Build your own prompt.
Anatomy of a prompt
A prompt generally has two parts, and being able to name them is worth a mark.
| Part | What it does | Example |
|---|---|---|
| Instruction | Tells the model what to do | Summarise this email in three bullets |
| Context | Gives the model the material it needs to do it | The email contains customer feedback from the past week |
A prompt that scores well is clear and concise, specific enough to steer the model, contextual enough to be answerable, and relevant to the task. Vague prompts produce vague output, and the fix is almost always a sharper instruction rather than a different model.
Inputs — the part that makes it reusable
An instruction with the text baked in is a one-off. An instruction with inputs is a function.
Inputs are placeholders you define in the prompt and fill at runtime. In prompt builder you add one by typing / or by selecting Add content, then choosing an input type.
| Input type | Carries | Typical source |
|---|---|---|
| Text | A string value | A column from a Dataverse row, a form field, a variable from a flow |
| Image or document | A file for the model to read | An attachment, a photo captured in an app, a file from a document library |
Why this matters for the exam
The distinction that gets tested is input versus knowledge. An input is data the caller hands over for this one run — the specific complaint, the specific invoice image. Knowledge is a body of organisational data the prompt can look into on every run. If the scenario says “the flow passes the email body to the prompt”, that is an input. If it says “the prompt should answer using our returns policy records”, that is knowledge.
Knowledge — grounding the answer in your data
A knowledge object represents a set of data retrieved through a data source connection, such as Dataverse. Attaching one lets the prompt generate answers that reflect your organisation’s own records rather than only what the model learned during training.
You add it in the Knowledge section of prompt builder, alongside the inputs.
Test before you save
Prompt builder gives you a test pane for a reason. You type a sample value for each input, select Test, and read the response.
What you are actually checking
- Does the shape match what the caller expects — three bullets when you asked for three bullets?
- Does it hold up on an awkward input as well as a tidy one — an empty field, a very long email, a complaint in another language?
- Is it stable across repeated runs, or does the wording drift each time?
Save only once the response looks right. Everything downstream inherits whatever you save.
Settings — tuning behaviour without rewriting the words
The settings panel opens from the three dots (…) → Settings at the top of prompt builder.
| Setting | Controls | When you change it |
|---|---|---|
| Temperature | How deterministic versus creative the answer is, on a 0–1 slider | Lower for accuracy and consistency, higher for drafting and idea generation |
| Record retrieval | How many records are pulled from your knowledge sources | Raise it when answers miss relevant records, lower it to keep responses focused |
| Include links in the response | Whether citations to the retrieved records appear in the answer | Turn on when a human reviewer needs to verify where an answer came from |
| Enable code interpreter | Whether the model may generate and execute code | Turn on for calculation or data-manipulation tasks |
| Content moderation level | How strictly harmful content is filtered — Low, Moderate, or High | Raise for stricter filtering; lower only with a genuine business reason |
Two behaviours here are worth remembering precisely, because they are the kind of detail an exam uses to separate people who have opened the panel from people who have only read about it:
Two settings that are not always available
- Temperature is unavailable for reasoning models. The slider is disabled when a reasoning model is selected — the model does not take the parameter.
- Content moderation level applies to managed models only. Choose an external model and the slider becomes unavailable, because moderation is applied by the hosting service rather than by the prompt.
Low temperature is not “worse” and high temperature is not “better”. A prompt that classifies a complaint into one of five categories wants 0. A prompt that drafts three alternative apology openings wants something higher. Match the setting to the job.
Choosing a model
The Model selector sits at the top of prompt builder. The dropdown lists the generative AI models available to generate answers for your prompt, and the list changes over time as models are added and updated.
Rather than memorising version numbers, learn the axes you are choosing along.
| Axis | What varies | How it affects your choice |
|---|---|---|
| Rate | Basic, standard, or premium consumption | A cheap model called a million times may cost more than a premium model called rarely |
| Context window | How much text the model can consider at once | Long documents need a larger window or the input gets truncated |
| Hosting | Managed models versus external models from other providers | External models carry that provider's terms and data handling, and lose the moderation slider |
| Region | Which models are offered where | A model available in one region may not be available in another |
Human oversight is part of the design
Generative models can be wrong and can carry bias. Microsoft’s guidance is explicit: a human should review generated content before it is published, sent to a customer, or used to inform a business decision.
Designing the review in, not bolting it on
In a flow, that usually means the prompt output goes to an approval rather than straight to a customer. In an app, it means the generated text lands in an editable field a person can correct before saving. If a scenario describes AI-generated content going directly to an external recipient with no review step, that is the flaw the question is testing.
Quick check
Nadia builds a prompt that must summarise a customer complaint the flow passes in, and must reference the company's published returns policy stored in a Dataverse table. How should she configure the prompt?
A maker selects a reasoning model for a prompt and notices the temperature slider is greyed out. What is the correct interpretation?
Which statement about starting a new prompt is correct?
What to carry forward
- A prompt is build once, call from anywhere — flows, apps, and agents share one definition.
- Prompts need Dataverse, a supported region, and credits.
- Start from a template for a standard task, build your own for a specific one.
- Inputs are per-run values from the caller. Knowledge is organisational data the prompt draws on every run.
- Temperature trades determinism for creativity, and is unavailable on reasoning models.
- Content moderation applies to managed models; external models lose the slider.
- Always test with awkward inputs before saving, and design a human review step into anything customer-facing.
Next you will take these prompts — and the prebuilt AI models alongside them — and wire them into apps and flows.